Quoting Sylvain Gelly <sylvain.ge...@m4x.org>:
On Fri, Jan 30, 2009 at 9:29 AM, Magnus Persson
<magnus.pers...@phmp.se>wrote:
Did you try to tune the bias constant at all or just took the one I posted?
I wrote it from memory and I believe it is in the range of possibly good
values, but it is certainly not optimal, I can be off by a significant
factor in both directions.
Yes I always tries to test at least 5 values. Three values centered
around what I currently believe is the best value and two or more
values to the extremes to get the big picture and make sure that I am
at least the maximum.
Constant Winrate Against Gnugo
0.00015 43.9
0.0015 50
0.015 50
0.15 43.1
1 7.1
10 6.7
The bias can't possibly be 1 because by definition it is the difference
between the expected value of the normal tree search and the expected value
of AMAF. As those two values are in [0,1] anyway, the bias is certainly not
1, and in most cases very small.
I think I was confused about exactly what in the term that was the
bias. I do not really remember what and why I thought these things.
Anyway your explanation is very clear.
On a side note I think Valkyria for some moves have very large biases
because the playouts are very heavy. I am thinking of an experiment to
test this. One way is to simply play uniform playouts from a single
position N times for all moves and thus get accurate win rates for all
those moves as well as AMAF values. Then taking the difference in
values for AMAF and real win rates should reveal the size of bias in
general and if there are systematic difference for certain patterns or
situations.
Alternatively one could play N playouts for a single move and see what
AMAF values one get. Now can see to what extent the biases are stable
when moves are played on the board. The real values will change for
all moves and the AMAF as well, but will the bias be constant? Or will
it be random as a function of the position.
There is no tree search in these two cases.
Selective search is often similar to the second situation, and my
question is whether this can cause larger systematic biases for some
moves, that thus is missed. And if one understands the nature of such
biases maybe they can be neutralized somehow.
It could be that using patterns to start with strong biases in the
prior AMAF values is doing exactly this.
Has anyone has done anything like this or have any interesting
opinions I would be thankful.
-Magnus
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
